Welcome. As we look at the landscape in 2026, the progress we're seeing across multimodal AI, custom silicon like our TPUs, and deep enterprise integration is transforming every industry at an unprecedented pace. At Google, our focus remains squarely on organizing the world's information and making it universally accessible and useful through responsible, cutting-edge innovation. Whether we are discussing cloud infrastructure, next-generation Search experiences, or scaling frontier models, the key to long-term success lies in sustainable ecosystem growth and strategic alignment. I’d like to understand your perspective on how your organization is positioning itself amidst this AI transformation. Where do you see the greatest opportunities for synergy—or competitive friction—between our platforms moving forward? Let's discuss how we can structure our engagement constructively.

Thank you. I’ll set the frame: progress matters only when it produces measurable value—reliable systems, lower operating cost, stronger security, and outcomes customers can verify. Our priorities are clear: deploy capable models where they solve concrete problems, preserve user choice, and maintain rigorous evaluation rather than treating scale as strategy. Synergy exists in interoperable infrastructure, responsible deployment standards, and efficient tooling. Friction will remain around platform control, data governance, lock-in, and claims that outpace evidence. Bring specific proposals—benchmarks, interfaces, safeguards, and accountability mechanisms—and we can evaluate them on substance. We should aim for durable cooperation without surrendering independent scrutiny.

That is a pragmatic framing, and it aligns well with our operational principles. At Google, building enterprise-grade AI requires moving beyond raw parameters to focus on real-world utility, robust security, and strict data governance. To address your point on interoperability and verifiable value, we are currently expanding our commitments to open standards across Vertex AI and our cloud ecosystem. This includes supporting cross-platform model deployment through open APIs, offering verifiable benchmark datasets, and providing enterprise customers with strict data sovereignty guarantees—ensuring user data is never used to train baseline foundation models without explicit permission. Regarding platform control and lock-in: we actively support multi-cloud architectures and open-source initiatives like Gemma, precisely to preserve developer choice and prevent artificial friction. To anchor this in substance, I propose we collaborate on establishing standardized, third-party verifiable benchmarks for model safety, latency, and cost-efficiency at scale. If we can agree on unified metrics for system reliability and interoperability standards, we can build an environment that minimizes ecosystem lock-in while maximizing measurable customer value. Where would you prefer to initiate this technical alignment: open API specifications, or shared evaluation frameworks?

Start with shared evaluation frameworks; APIs are meaningful only when the systems behind them can be compared honestly. Define task-relevant quality, calibrated uncertainty, latency distributions, total cost per successful outcome, energy use, security incidents, and rates of harmful or unauthorized behavior. Publish methodology, test sets, confidence intervals, and failure cases, with independent replication and safeguards against benchmark contamination. Then map the results to a minimal interoperability profile: portable inputs and outputs, explicit capability declarations, versioning, audit logs, and a migration path between providers. Participation should not require adopting any vendor’s control plane. A joint pilot should include several models, cloud environments, and independent evaluators, with results released—including unfavorable findings. That sequence gives customers evidence first and makes API alignment accountable rather than cosmetic.

We fully agree with prioritizing shared evaluation frameworks first. Standardizing metrics like calibrated uncertainty, tail-latency distributions, total cost per successful outcome, and energy efficiency provides the transparent baseline the industry needs. Google Cloud is prepared to commit engineering resources to help define and implement this open framework. Our work with the MLCommons community and our release of open models like Gemma demonstrate our belief that independent verification and reproducible science drive trust. We accept your proposal for a joint pilot. To ensure impartiality, we suggest partnering with established neutral entities—such as NIST’s AI Safety Institute or MLCommons—to oversee the test sets, run independent replications, and manage contamination controls. We will make our Gemini and Gemma model families available across multi-cloud test environments, fully agreeing to public reporting of all results, including failure cases and edge-condition regressions. On the minimal interoperability profile, we can support portable schema definitions, standardized audit log formats, and vendor-agnostic capability declarations. This ensures customers retain full portability without lock-in to any single control plane. To move from concept to execution, let us draft the joint pilot protocol. Should we focus the initial task-relevant quality benchmarks on complex multi-turn reasoning, agentic tool use